contextforge-mcp

MCP compression middleware that connects codebase-memory-mcp + headroom + Spec Kit into a unified, token-efficient workflow.

Architecture

Claude Code
  ├── codebase-memory-mcp   ← graph queries (cbm_* tools)
  │         │
  │         └── large result
  │                   │
  └── contextforge-mcp  ← YOU ARE HERE
            │
            cf_compress_cbm(result, tool_name)
            │
            └── compressed result (60-95% fewer tokens)

ContextForge does NOT proxy codebase-memory-mcp — both servers run independently. The agent calls CBM for graph queries, then passes results through ContextForge for compression. This design is reliable, cross-platform, and works with any MCP client.

Install

npm install -g codebase-memory-mcp
pip install "headroom-ai[all]"
pip install contextforge-mcp

Setup

# Health check
contextforge-mcp doctor

# Configure Claude Code (writes .mcp.json with both servers)
contextforge-mcp install --target claude

Workflow

# 1. Query the graph (via codebase-memory-mcp)
result = cbm_search_graph(name_pattern=".*Payment.*", label="Function")

# 2. Compress the result (via contextforge-mcp)
compressed = cf_compress_cbm(result=result, tool_name="search_graph")
# → [ContextForge ✓ search_graph: 8420→612 tokens (93% saved in 45ms)]

# 3. Use compressed result in your context
# 4. Check savings
cf_stats()

Tools (9 total)

Compression

Stats

Spec Kit

Supported CBM tool names for cf_compress_cbm

search_graph · search_code · get_architecture · find_dead_code · find_similar_code · get_impact · trace_path · trace_call_path · cypher_query · get_cross_service_links · get_node_details

Add to CLAUDE.md

## ContextForge MCP — Compression Workflow

After EVERY codebase-memory-mcp tool call that returns a large result,
immediately call cf_compress_cbm(result, tool_name) to compress it.

| CBM Query | Then compress with |
|-----------|-------------------|
| cbm_search_graph(…) | cf_compress_cbm(result, "search_graph") |
| cbm_get_architecture() | cf_compress_cbm(result, "get_architecture") |
| cbm_search_code(…) | cf_compress_cbm(result, "search_code") |
| cbm_trace_path(…) | cf_compress_cbm(result, "trace_path") |
| cbm_get_impact(…) | cf_compress_cbm(result, "get_impact") |

Call cf_stats() at end of session to measure total savings.

Credits

• codebase-memory-mcp — MIT
• headroom — Apache 2.0
• spec-kit — MIT

License

MIT